Interpretable Machine Learning for Exchange Rate Forecasting with Macroeconomic Fundamentals

2.1 Data & Variables

A set of macroeconomic and financial variables hypothesized to influence the CAD/USD rate was collected. This likely includes:

• Commodity Prices: Crude oil (WTI), gold, natural gas.
• Financial Indicators: S&P/TSX Composite Index, interest rate differentials (Canada vs. U.S.).
• Macroeconomic Fundamentals: GDP growth, inflation differentials, trade balance.

Data is preprocessed (e.g., stationarity transformations, handling missing values) to suit ML models.

2.2 Machine Learning Models

The study likely utilizes powerful, yet complex, ensemble models known for high predictive accuracy:

• Gradient Boosting Machines (GBM/XGBoost/LightGBM): Effective for capturing nonlinear patterns and interactions.
• Random Forests: Robust to overfitting and provides inherent feature importance measures.
• Neural Networks: Potentially used for capturing deep, complex temporal dependencies.

Models are trained to predict future exchange rate movements or levels.

2.3 Interpretability Techniques

To open the "black box," the study applies state-of-the-art IML methods:

• SHAP (SHapley Additive exPlanations): A game-theoretic approach to quantify the contribution of each feature to each individual prediction. It provides both global and local interpretability.
• Partial Dependence Plots (PDPs): Visualize the marginal effect of a feature on the predicted outcome.
• Feature Importance Rankings: Derived from model-specific metrics or permutation importance.

These techniques help answer *why* a certain prediction was made.

3.1 Model Performance

The machine learning models demonstrated superior predictive accuracy compared to traditional linear benchmarks (e.g., Vector Autoregression - VAR). Performance was evaluated using metrics like Root Mean Squared Error (RMSE), Mean Absolute Error (MAE), and possibly directional accuracy. The results validate the capability of ML to model complex exchange rate dynamics.

3.2 Feature Importance & SHAP Analysis

The interpretability analysis yielded clear, economically intuitive insights:

1. Crude Oil Price: Emerged as the most significant determinant. SHAP values revealed its effect is time-varying, with changes in sign and magnitude aligning with major events in commodity markets (e.g., the 2014 oil price crash, OPEC+ decisions). This aligns with Canada's evolving oil export landscape.
2. Gold Price: The second most important variable, acting as a safe-haven asset and inflation hedge influencing the CAD.
3. TSX Stock Index: Ranked third, reflecting the domestic economic health and capital flows.

Chart Description (Implied): A SHAP summary plot would show each variable as a row. For crude oil, dots would be spread across both positive and negative SHAP values on the x-axis (impact on prediction), with color indicating the feature's value (e.g., blue for low oil price, red for high). This visually confirms the time-varying and non-monotonic relationship.

3.3 Ablation Study for Model Refinement

A key innovation is using interpretation outputs (like low-importance features identified by SHAP) to guide an ablation study. Features deemed less important are iteratively removed, and model performance is re-evaluated. This process:

• Simplifies the model, reducing overfitting and computational cost.
• Potentially improves predictive accuracy by eliminating noise.
• Creates a more parsimonious and focused final model, enhancing practical utility.

5.1 Mathematical Formulation

The core predictive model can be represented as:

$\hat{y}_t = f(\mathbf{x}_t) + \epsilon_t$

where $\hat{y}_t$ is the forecasted exchange rate return or level at time $t$, $f(\cdot)$ is the complex function learned by the ML model (e.g., a gradient boosting ensemble), $\mathbf{x}_t$ is the vector of input features (oil price, gold, TSX, etc.), and $\epsilon_t$ is the error term.

The SHAP value $\phi_i$ for feature $i$ for a single prediction explains the deviation from the average prediction:

$f(\mathbf{x}) = \phi_0 + \sum_{i=1}^{M} \phi_i$

where $\phi_0$ is the base value (average model output) and $M$ is the number of features. $\phi_i$ is calculated using the classic Shapley value formula from cooperative game theory, considering all possible feature combinations:

$\phi_i = \sum_{S \subseteq \{1,\ldots,M\} \setminus \{i\}} \frac{|S|! \, (M - |S| - 1)!}{M!} [f_{S \cup \{i\}}(\mathbf{x}_{S \cup \{i\}}) - f_S(\mathbf{x}_S)]$

This ensures a fair attribution of the prediction to each feature.

5.2 Analysis Framework Example

Scenario: Understanding the model's prediction for a strong CAD appreciation on a specific date.

Step-by-Step IML Analysis:

1. Local SHAP Explanation: Generate force plot or waterfall plot for the specific prediction.
   • Output: "Prediction: CAD appreciates by 1.5%. Key drivers: WTI Oil (+1.1%), Gold Price (+0.3%), TSX (-0.2% due to a slight drop)."
2. Contextual Check: Cross-reference with market events.
   • Action: "On this date, OPEC+ announced a production cut, spiking oil prices. The model's high positive SHAP for oil aligns perfectly with this fundamental shock."
3. PDP Analysis: Examine the PDP for oil prices.
   • Observation: "The PDP shows a steep positive slope at current price levels, confirming the model is in a regime where oil price increases strongly boost the CAD."
4. Ablation Feedback: If, for many predictions, a feature like "U.S. Industrial Production" has near-zero SHAP values, it becomes a candidate for removal in the next model training iteration to enhance simplicity and robustness.